All Search Is Local Search: Positioning, Context & Measurement

Location, location, location.

I mentioned in my opening post on this site how I’ve spent a large majority of my time working with small businesses and local search.

For most small businesses local search is the most relevant flavor of search, as their market is largely dictated by their location; that is, their customers are within some radius of their main place of business.

Understanding customers’ preference for local businesses in many categories and verticals, search engines align their search results to position those small businesses closer to them in search in relevant moments, for relevant queries/prompts.

If you look closely at my wording there, it’s easy to generalize (something we’re programmed to do in the mathematics world) that statement and land on an important result.

All search is local search.

No matter what kind of content you put online for an audience, it’s going to be digested and end up occupying some position on (or in) different search surfaces. Whether they’re addressable memory spaces (landing in more fixed, static regions) or associative memory surfaces (landing in more fluid, dynamic regions), what you add to digital search spaces must occupy some point (or points) in that space.

Occupying a point (or points) in a search space means it has some location (static or fluid), relative to other items in that space. That means for every search, every query, every prompt, there’s some related region or area in that space that is most relevant to that search, query, prompt, et. al. This region could be a topic, a business vertical, a product or service, a definition (and so on) — whatever is most relevant to that search in that particular moment.

The goal for the search engine (AI/LLM search or otherwise), then, is to chart a function that maps those searches/prompts/queries to those specific regions in its search space to retrieve candidates that occupy that region – ultimately putting those results closer to the user during those moments of search (and making the decision mechanism on choosing results/responses more efficient).

Each search/prompt/query, then, is mapped to some local region (or neighborhood): making every search a local search within a representational candidate space, essentially.

Positioning In Representational Candidate Spaces

I write extensively in this blog about the mathematical structure and nuances of the learned representational spaces inside search/AI/LLMs.

Ultimately digital content is ingested into those spaces through well known and studied mathematical machinery ( see my walkthrough on topology, metric spaces, vector spaces and inner product spaces for more details if you’d like ), making it easier for the architectures to work with – including what regions your content is most relevant to, in different contextual situations.

In more traditional search spaces, those locations were a bit more static; that is, content (web pages, et. al.) generally held a more definite, relative position in those spaces for some subset of queries of the query space.

This rigidity, while reliable and stable, meant the search space wasn’t as flexible for more nuanced searches — ones that required more understanding of the meaning or intent behind the search; thus the move to more associative memory spaces, largely served by vector space models that could overlay the static search space with a space that served meaning instead of raw dictionary-like lookups.

The move to a more “soft” lookup in more granular associative spaces — like the ones found in modern search/AI/LLM spaces – meant that positioning for web pages, passages and other items in those spaces became more fluid over time.

Regardless of the evolution of the representational space, you will still hold a position in that space — a position that is meaningful to some subset of queries, prompts and conversations related to that particular region in space.

Context & Attention

As mentioned above, the fluidity of associative spaces gives it more flexibility in different search/prompt situations – flexibility dictated by context and attention.

The inherent attention mechanisms in modern search/AI/LLMs are used to understand what’s most relevant during the course of a search/prompt session — and derive what region (or regions) will most likely have the best starting points (and set of candidates) for retrieving/assembling a result or response.

As more context is added to the search/prompt, those regions in space can change and refine the starting point for the eventual set of results or responses, based on the related context.

“track shoes” might have a broad starting point and dense (competitive) candidate space vs “track shoes for 400 meter runners”, which will likely have a much more specific starting point and less dense area of the representational space.

Other contextual elements might help resolve that starting point even further and shape the output – if a user’s location or brand preference is known from some other learned activity, for example.

Measurement Walks In AI/LLM Search Spaces

So, how do we ultimately measure these more fluid spaces – and the moving targets that are the starting points in these spaces?

I’ll be writing much more on measurement soon, but as a placeholder I’d like to offer a mental model that’s related to the “local” thread above.

I’ve mentioned in a previous post, that measurement will ultimately be defined by some contextual basis – potentially taking elements of user preferences, user activity, location, device, network – among other types of context used to produce results/responses.

Subtle changes in context (even wording) could map you to different starting points in regions of those spaces. Every prompt is, in effect, a measurement on the space itself, giving you clues to that inherent region of space for different observers or observations.

The recursive nature of the responses in AI/LLM spaces can yield a new “walk” through the space each time – thinking of each word as a step on a path a map.

Every query or prompt (and related context) can collapse the possible walks (from a particular starting point or region) on that map to a particular subset or set of branches on a path, depending on context.

Some paths or walks are more deterministic (like definitions or consensus searches) – some others have more possible branches to choose from, depending on regional starting point, candidates and decoding policy.

Given some initial starting point (location) a superposition of path possibilities exist, constrained by context form the different branches to choose from (selecting the next words, words, passages, etc.).

In a sense, we’re trying to make our websites, content, entities, brands, etc. present and contextually relevant to the most important walks or paths through these spaces – making sure we’re increasing the probability of being a “local step” on these paths for a particular user, in a particular moment.

Again, local.

Local Is Thematic, Not A Specific Flavor Of Search

While I’ve worked in many flavors of search over the years, they’re all really local search in some, way, shape or form.

Local to a geographic marker.

Local to topic.

Local to an industry/vertical.

Local to a product/service.

Local to an idea or concept.

Local to an intent.

However you slice the search space, you’re competing to be closer to your customers through the search space in those moments – and that means inclusion in local regions within those spaces that folks will have to traverse, mapped from a particular covering of relevant prompts or queries.

Location. Location. Location.